Apple M5 for Local AI: MacBook Pro M5, M5 Pro & M5 Max (128GB) Guide

Short answer: if you want a Mac for serious local AI in 2026, buy the MacBook Pro M5 Max with the 40-core GPU and 128GB of unified memory (614GB/s, from $3,899 before the memory upgrade). Since Apple cut the Mac Studio's ceiling to 96GB during the memory shortage, the M5 Max is the only 128GB Mac you can buy. Expect ~15-18 tok/s on dense 70B models and ~78-82 tok/s on Qwen's 32B MoE models per the llmcheck.net leaderboard. The base M5 (153GB/s, 32GB max) is for 8B-class models only.

That is the whole recommendation. The rest of this guide is the evidence: what each M5 tier actually costs and runs, published tokens-per-second numbers with sources, the setup commands, and the honest list of things a $4,000+ laptop still does worse than a $2,000 GPU. If you are choosing between Mac generations or between Mac and PC entirely, the Apple Silicon buying guide and Mac Studio vs PC comparison cover those calls.

---

The M5 Lineup, in Local-AI Terms {#m5-lineup}

The spec that matters is memory bandwidth, because token generation is bandwidth-bound — and the M5 tiers are 153, 307, and 460/614GB/s. Everything else on the spec sheet is secondary for LLM work.

All figures below are from Apple's published specs and the March 3, 2026 announcement (M5 Pro/Max became available March 11; the base M5 MacBook Pro shipped October 2025):

Three things worth knowing beyond the table:

• Every GPU core carries a Neural Accelerator in this generation — Apple's claim is "up to 4x faster LLM prompt processing than M4 Pro and M4 Max." That is a prefill claim, not a generation-speed claim; we unpack the difference in the benchmark section.
• Base storage doubled: M5 Pro models start at 1TB, M5 Max at 2TB. Model files are huge — a Q4 70B is ~40GB and you will hoard several — so this matters more than it sounds.
• You do not need the top storage tiers. 128GB of memory plus the stock 2TB is the sane AI build. Apple has moved memory pricing and availability more than once this year — it deleted Mac Studio tiers outright mid-cycle — so check the configurator for the current cost of the 128GB upgrade rather than trusting any number published even a month ago.

If you are deciding how much memory a specific model needs before you spec anything, our Apple Silicon AI calculator does the arithmetic for you.

---

Why the M5 Max Is Now the Only 128GB Mac {#only-128gb-mac}

As of August 2026, Apple's own configurator tops the Mac Studio out at 96GB — so the M5 Max MacBook Pro at 128GB is the highest-memory Mac Apple sells. There is no M5 Ultra, and the M3 Ultra's big-memory tiers are gone.

This is the strangest Apple lineup situation in years, and it happened in two steps:

1. The 512GB Mac Studio option disappeared in March 2026. MacRumors reports Apple quietly pulled the M3 Ultra's 512GB memory tier amid rising RAM prices — the same memory squeeze that has GPU street prices 50%+ over MSRP.
2. The 256GB and 128GB tiers followed by May. Apple's current Mac Studio specs page lists the M4 Max at 36GB or 64GB and the M3 Ultra fixed at 96GB, not configurable. No 128GB, no 256GB, no 512GB — checked against Apple's live specs page in August 2026.

So the machine Apple positions as its studio workstation now carries less memory than its laptop. If your local-AI plan is "128GB of unified memory," the M5 Max MacBook Pro is not the best Apple option — it is the only Apple option. (The M3 Ultra Mac Studio at 96GB is still interesting: 819GB/s of bandwidth beats the M5 Max's 614GB/s. More on that trade in the alternatives section.)

And no, there is no M5 Ultra. The Mac Studio has not moved to the M5 generation at all. Whether Apple restores the big-memory Studio tiers when DRAM supply normalizes is anyone's guess; buying decisions have to be made against the configurator that exists today.

---

Real Benchmarks: What an M5 Max Actually Generates {#benchmarks}

On a 128GB M5 Max, published leaderboard runs show dense 70B models at 15-18 tok/s (Q4_K_M, MLX), 32B-class Qwen models at ~80 tok/s, and 8B models at ~138 tok/s. Attribution: llmcheck.net's Apple Silicon leaderboard, which aggregates 227 benchmark data points across Ollama, LM Studio, and MLX.

Source: llmcheck.net/benchmarks, MLX runtime, 128GB M5 Max, August 2026. Type labels on the Qwen 4 rows are omitted because the leaderboard does not publish them — and the throughput suggests low-activation MoE rather than dense. *The 235B row is listed at Q4_K_M by the leaderboard, but 235B of weights at Q4 is ~135-140GB — more than 128GB of RAM — so the run was almost certainly a ~3-bit quant; we relabel it accordingly. Runtime matters too: the same leaderboard lists DeepSeek R1 70B at 11 tok/s under Ollama, so treat 11-18 tok/s as the honest 70B band.

How to read that table honestly:

• MoE models are the sweet spot. Qwen's 32B-A3B runs at 82 tok/s because only ~3B parameters activate per token — you get 32B-class quality at small-model speed. Even the 235B-A22B monster produces 18 tok/s, the same as a dense 70B, while using far more memory. On a 128GB machine, MoE models are what the capacity is for.
• Dense 70B at 15-18 tok/s is reading speed, not build speed. Fine for chat, agents, and code review; painful for batch jobs.

Versus the M4 Max: on the directly comparable leaderboard row (Llama 5 70B, MLX, Q4), the M5 Max generates 18 tok/s to the M4 Max's 15 — about 20% faster, in line with the bandwidth bump from 546 to 614GB/s. The bigger generational jump is prefill: Apple's claim for this generation is up to 4x faster LLM prompt processing than M4 Pro and M4 Max, credited to the Neural Accelerator in each GPU core — a vendor number until third parties re-test it, but it targets exactly the workload the accelerators were built for. In practice that means long-context work — pasting a codebase, RAG over big documents — is where M5 feels like a new machine, while raw generation speed improved only modestly. If you already own an M4 Max with 128GB, stay put; this is not an upgrade generation for you.

---

Which Config to Buy {#which-config}

Buy the memory tier your target models need, then stop: 32GB for 8B-14B models, 64GB (M5 Pro or Max) for 32B-class, 128GB (M5 Max only) for 70B+ and big MoE. Overspending on CPU cores or storage does nothing for tokens per second.

16GB / 24GB — the accidental AI Mac. If you have (or buy) a base-M5 machine for other reasons, it runs 4B-8B models genuinely well — figure roughly 30-35 tok/s on an 8B at Q4 given the base chip's 153GB/s bandwidth (the 138 tok/s in the table above is the 614GB/s M5 Max; generation scales with bandwidth). That is still faster than most people read code. Our best LLMs for a 16GB Mac page ranks exactly what fits in that footprint.

32GB — the honest budget pick ($1,699-and-up territory, base M5). A Q4 14B fits with room for context; a Q4 32B technically loads but the 153GB/s bandwidth makes it a slideshow. The model picks for this tier are on our best LLMs for a 32GB Mac page.

48GB / 64GB M5 Pro ($2,199+). 307GB/s doubles the base chip's bandwidth. The one published 64GB M5 Pro leaderboard row — Qwen 4.1 32B-A3B at 56 tok/s under MLX (llmcheck.net) — is a genuine daily-driver number for a coding model, and since generation scales with bandwidth, expect dense models at roughly half the M5 Max figures above. A Q4 70B (~40GB) squeezes into 64GB but budget your patience.

128GB M5 Max — the reason this page exists. Take the 40-core GPU (614GB/s; the 32-core variant is 460GB/s and gives up a quarter of your future tokens), take 128GB, keep the stock 2TB. 14-inch from $3,599 or 16-inch from $3,899 before the memory upgrade — the 16-inch has more thermal headroom for sustained inference, which is what LLM serving is.

One warning from this year's pattern: Apple has already removed memory tiers from the Mac mini and Mac Studio mid-cycle during the DRAM crunch. If the 128GB MacBook Pro tier is the machine you want, there is a non-zero risk that configuration gets more expensive or constrained rather than cheaper.

---

Setup: Ollama and MLX on an M5 {#setup}

Two runtimes cover everything: Ollama for convenience and model breadth, MLX for Apple-optimized speed. Both are free; install both.

Ollama (easiest start — models in one command):

# Install (or download the app from ollama.com/download/mac)
brew install ollama

# Pick by memory tier
ollama run llama3.1:8b     # 16GB+ — fast on any M5
ollama run qwen3.6:27b     # 32GB+ — best small dense model
ollama run gemma3:27b      # 32GB+ — multimodal, 128K context
ollama run llama3.3:70b    # 64GB works, 128GB comfortable

MLX (Apple's own framework — typically 5-15% faster than Ollama on the same model, per llmcheck's cross-runtime data):

pip install mlx-lm

# The mlx-community org on Hugging Face maintains Mac-ready conversions
mlx_lm.generate --model mlx-community/Llama-3.3-70B-Instruct-4bit \
  --prompt "Explain unified memory in one paragraph"

# Interactive chat
mlx_lm.chat --model mlx-community/Llama-3.3-70B-Instruct-4bit

Note the leaderboard numbers above were measured under MLX — it is the faster path for big dense models on M5, and the Neural Accelerators specifically speed up its prefill.

Two practical tips for a 128GB machine:

• macOS caps GPU-wired memory below your total RAM (roughly 75% by default). The community-documented override is sudo sysctl iogpu.wired_limit_mb=110000 to let a 128GB Mac dedicate ~110GB to models. It resets on reboot; use it when a big MoE almost fits.
• Watch memory pressure, not free GB. Activity Monitor's memory pressure graph should stay green while a model is loaded. Yellow means swapping, and swapping murders tokens per second.

Full first-hour walkthrough — models, context settings, the works — in our Llama on Mac setup guide.

---

Honest Limitations {#limitations}

A 5090-class GPU still generates 2-3x faster on anything that fits in its VRAM, CUDA-only tooling still does not run, and Apple's "4x faster" AI claim is about prefill — not generation. Know what you are not buying:

• Generation speed is a bandwidth story Apple only partly fixed. 614GB/s is excellent for a laptop and roughly a third of a flagship GPU's GDDR7. Dense 70B at 15-18 tok/s is the ceiling until Apple ships more bandwidth. If your models fit in 24-32GB of VRAM, a discrete GPU is simply the faster tool.
• The 4x prompt-processing claim is real but narrow. Prefill got dramatically faster thanks to the per-core Neural Accelerators; sustained token generation improved ~20% generation-over-generation on the comparable leaderboard row. Reviews that only quote time-to-first-token flatter the machine.
• No CUDA. Fine-tuning tooling on Mac (MLX-LM does QLoRA) is workable for 7B-14B but years behind the CUDA ecosystem. Training seriously? Rent NVIDIA.
• The AI build is expensive and the price direction is up. $3,899 is the doorway, not the price — the 128GB config lands north of that, and the memory shortage has so far removed Apple's cheap-memory options rather than added them.
• Thermals are real on the 14-inch. Sustained inference is a space heater workload; the 16-inch chassis holds clocks better. Neither throttles catastrophically, but do not expect leaderboard numbers on battery in a quiet room — plugged-in, high-power mode is where those runs happen.

---

M5 Max vs the Alternatives {#vs-alternatives}

Buy the M5 Max for portable capacity; buy the M3 Ultra Studio for bandwidth under 96GB; buy a GPU tower for raw speed under 32GB. The one-table version:

The interesting fight is M5 Max vs M3 Ultra: the Studio's 819GB/s makes it the faster Mac for anything under its 96GB ceiling, and a dense 70B fits under 96GB with room to spare. Choose the Studio if it lives on a desk and 96GB is enough forever; choose the M5 Max if you want the 128GB headroom for the big MoE wave — Qwen3-235B-A22B at ~3-bit (~100GB of weights) is exactly the kind of model that fits in 128GB but not 96GB — or a machine that closes and goes in a bag. The full desk-vs-tower argument is in Mac Studio vs PC for AI.

Against the PC, the math has not changed since the M4 generation: NVIDIA wins tokens per second, Apple wins model size per watt and per decibel. What changed in 2026 is the context — with GPU prices 50%+ over MSRP, the premium for going Mac shrank without Apple lowering a single price.

---

Verdict {#verdict}

The M5 Max 128GB is the best local-AI Mac ever shipped, and — for now — the only high-memory Mac at all. The decisions, compressed:

1. Serious about big local models on macOS? 16-inch M5 Max, 40-core GPU, 128GB, stock 2TB. From $3,899 plus the memory upgrade. It is the only 128GB Mac; there is no configurator trick around that.
2. Daily driver is a 32B-class coder? M5 Pro 64GB at $2,199-$2,699 delivers ~55 tok/s on the 32B-class MoE coders (per llmcheck.net) and saves well over a thousand dollars.
3. 8B models and curiosity? Any base M5 — even 16GB runs an 8B at speed. Spec 32GB if you can and check the 32GB Mac model picks.
4. Own an M4 Max 128GB? Skip this generation. +20% generation speed is not an upgrade; the prefill gains alone rarely justify $4K+.
5. Need raw speed on models under 32GB? A GPU tower still wins by 2-3x. Different tool, different job.

The memory shortage did something strange to this lineup: it turned a laptop into Apple's flagship AI machine by subtraction. Until a Studio-class M5 Ultra restores the big-memory desktop, the MacBook Pro M5 Max 128GB is where Apple local AI tops out — and it is a genuinely good place to top out.

---

Sources {#sources}

• Apple Newsroom — MacBook Pro with M5 Pro and M5 Max — March 3, 2026 announcement: prices, memory tiers, bandwidth, AI claims
• Apple — MacBook Pro tech specs and Mac Studio tech specs — current chip, memory, and bandwidth configurations (checked August 2026)
• llmcheck.net/benchmarks — Apple Silicon tokens-per-second leaderboard (227 benchmark data points across Ollama, LM Studio, and MLX; results averaged over 3 runs on a freshly booted system)
• MacRumors — Mac Studio roundup — the 512GB/256GB/128GB memory-tier removals amid rising 2026 RAM prices

---

FAQ {#faq}